2 * Block multiqueue core code
4 * Copyright (C) 2013-2014 Jens Axboe
5 * Copyright (C) 2013-2014 Christoph Hellwig
7 #include <linux/kernel.h>
8 #include <linux/module.h>
9 #include <linux/backing-dev.h>
10 #include <linux/bio.h>
11 #include <linux/blkdev.h>
12 #include <linux/kmemleak.h>
14 #include <linux/init.h>
15 #include <linux/slab.h>
16 #include <linux/workqueue.h>
17 #include <linux/smp.h>
18 #include <linux/llist.h>
19 #include <linux/list_sort.h>
20 #include <linux/cpu.h>
21 #include <linux/cache.h>
22 #include <linux/sched/sysctl.h>
23 #include <linux/delay.h>
24 #include <linux/crash_dump.h>
25 #include <linux/prefetch.h>
27 #include <trace/events/block.h>
29 #include <linux/blk-mq.h>
32 #include "blk-mq-tag.h"
34 static DEFINE_MUTEX(all_q_mutex);
35 static LIST_HEAD(all_q_list);
37 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx);
40 * Check if any of the ctx's have pending work in this hardware queue
42 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
44 return sbitmap_any_bit_set(&hctx->ctx_map);
48 * Mark this ctx as having pending work in this hardware queue
50 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
51 struct blk_mq_ctx *ctx)
53 if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw))
54 sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw);
57 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
58 struct blk_mq_ctx *ctx)
60 sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw);
63 void blk_mq_freeze_queue_start(struct request_queue *q)
67 freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
68 if (freeze_depth == 1) {
69 percpu_ref_kill(&q->q_usage_counter);
70 blk_mq_run_hw_queues(q, false);
73 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start);
75 static void blk_mq_freeze_queue_wait(struct request_queue *q)
77 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
81 * Guarantee no request is in use, so we can change any data structure of
82 * the queue afterward.
84 void blk_freeze_queue(struct request_queue *q)
87 * In the !blk_mq case we are only calling this to kill the
88 * q_usage_counter, otherwise this increases the freeze depth
89 * and waits for it to return to zero. For this reason there is
90 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
91 * exported to drivers as the only user for unfreeze is blk_mq.
93 blk_mq_freeze_queue_start(q);
94 blk_mq_freeze_queue_wait(q);
97 void blk_mq_freeze_queue(struct request_queue *q)
100 * ...just an alias to keep freeze and unfreeze actions balanced
101 * in the blk_mq_* namespace
105 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
107 void blk_mq_unfreeze_queue(struct request_queue *q)
111 freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
112 WARN_ON_ONCE(freeze_depth < 0);
114 percpu_ref_reinit(&q->q_usage_counter);
115 wake_up_all(&q->mq_freeze_wq);
118 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
120 void blk_mq_wake_waiters(struct request_queue *q)
122 struct blk_mq_hw_ctx *hctx;
125 queue_for_each_hw_ctx(q, hctx, i)
126 if (blk_mq_hw_queue_mapped(hctx))
127 blk_mq_tag_wakeup_all(hctx->tags, true);
130 * If we are called because the queue has now been marked as
131 * dying, we need to ensure that processes currently waiting on
132 * the queue are notified as well.
134 wake_up_all(&q->mq_freeze_wq);
137 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
139 return blk_mq_has_free_tags(hctx->tags);
141 EXPORT_SYMBOL(blk_mq_can_queue);
143 static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
144 struct request *rq, int op,
145 unsigned int op_flags)
147 if (blk_queue_io_stat(q))
148 op_flags |= REQ_IO_STAT;
150 INIT_LIST_HEAD(&rq->queuelist);
151 /* csd/requeue_work/fifo_time is initialized before use */
154 req_set_op_attrs(rq, op, op_flags);
155 /* do not touch atomic flags, it needs atomic ops against the timer */
157 INIT_HLIST_NODE(&rq->hash);
158 RB_CLEAR_NODE(&rq->rb_node);
161 rq->start_time = jiffies;
162 #ifdef CONFIG_BLK_CGROUP
164 set_start_time_ns(rq);
165 rq->io_start_time_ns = 0;
167 rq->nr_phys_segments = 0;
168 #if defined(CONFIG_BLK_DEV_INTEGRITY)
169 rq->nr_integrity_segments = 0;
172 /* tag was already set */
182 INIT_LIST_HEAD(&rq->timeout_list);
186 rq->end_io_data = NULL;
189 ctx->rq_dispatched[rw_is_sync(op, op_flags)]++;
192 static struct request *
193 __blk_mq_alloc_request(struct blk_mq_alloc_data *data, int op, int op_flags)
198 tag = blk_mq_get_tag(data);
199 if (tag != BLK_MQ_TAG_FAIL) {
200 rq = data->hctx->tags->rqs[tag];
202 if (blk_mq_tag_busy(data->hctx)) {
203 rq->cmd_flags = REQ_MQ_INFLIGHT;
204 atomic_inc(&data->hctx->nr_active);
208 blk_mq_rq_ctx_init(data->q, data->ctx, rq, op, op_flags);
215 struct request *blk_mq_alloc_request(struct request_queue *q, int rw,
218 struct blk_mq_ctx *ctx;
219 struct blk_mq_hw_ctx *hctx;
221 struct blk_mq_alloc_data alloc_data;
224 ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT);
228 ctx = blk_mq_get_ctx(q);
229 hctx = q->mq_ops->map_queue(q, ctx->cpu);
230 blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx);
232 rq = __blk_mq_alloc_request(&alloc_data, rw, 0);
233 if (!rq && !(flags & BLK_MQ_REQ_NOWAIT)) {
234 __blk_mq_run_hw_queue(hctx);
237 ctx = blk_mq_get_ctx(q);
238 hctx = q->mq_ops->map_queue(q, ctx->cpu);
239 blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx);
240 rq = __blk_mq_alloc_request(&alloc_data, rw, 0);
241 ctx = alloc_data.ctx;
246 return ERR_PTR(-EWOULDBLOCK);
250 rq->__sector = (sector_t) -1;
251 rq->bio = rq->biotail = NULL;
254 EXPORT_SYMBOL(blk_mq_alloc_request);
256 struct request *blk_mq_alloc_request_hctx(struct request_queue *q, int rw,
257 unsigned int flags, unsigned int hctx_idx)
259 struct blk_mq_hw_ctx *hctx;
260 struct blk_mq_ctx *ctx;
262 struct blk_mq_alloc_data alloc_data;
266 * If the tag allocator sleeps we could get an allocation for a
267 * different hardware context. No need to complicate the low level
268 * allocator for this for the rare use case of a command tied to
271 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
272 return ERR_PTR(-EINVAL);
274 if (hctx_idx >= q->nr_hw_queues)
275 return ERR_PTR(-EIO);
277 ret = blk_queue_enter(q, true);
281 hctx = q->queue_hw_ctx[hctx_idx];
282 ctx = __blk_mq_get_ctx(q, cpumask_first(hctx->cpumask));
284 blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx);
285 rq = __blk_mq_alloc_request(&alloc_data, rw, 0);
288 return ERR_PTR(-EWOULDBLOCK);
293 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
295 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
296 struct blk_mq_ctx *ctx, struct request *rq)
298 const int tag = rq->tag;
299 struct request_queue *q = rq->q;
301 if (rq->cmd_flags & REQ_MQ_INFLIGHT)
302 atomic_dec(&hctx->nr_active);
305 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
306 blk_mq_put_tag(hctx, ctx, tag);
310 void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx, struct request *rq)
312 struct blk_mq_ctx *ctx = rq->mq_ctx;
314 ctx->rq_completed[rq_is_sync(rq)]++;
315 __blk_mq_free_request(hctx, ctx, rq);
318 EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request);
320 void blk_mq_free_request(struct request *rq)
322 struct blk_mq_hw_ctx *hctx;
323 struct request_queue *q = rq->q;
325 hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu);
326 blk_mq_free_hctx_request(hctx, rq);
328 EXPORT_SYMBOL_GPL(blk_mq_free_request);
330 inline void __blk_mq_end_request(struct request *rq, int error)
332 blk_account_io_done(rq);
335 rq->end_io(rq, error);
337 if (unlikely(blk_bidi_rq(rq)))
338 blk_mq_free_request(rq->next_rq);
339 blk_mq_free_request(rq);
342 EXPORT_SYMBOL(__blk_mq_end_request);
344 void blk_mq_end_request(struct request *rq, int error)
346 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
348 __blk_mq_end_request(rq, error);
350 EXPORT_SYMBOL(blk_mq_end_request);
352 static void __blk_mq_complete_request_remote(void *data)
354 struct request *rq = data;
356 rq->q->softirq_done_fn(rq);
359 static void blk_mq_ipi_complete_request(struct request *rq)
361 struct blk_mq_ctx *ctx = rq->mq_ctx;
365 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
366 rq->q->softirq_done_fn(rq);
371 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
372 shared = cpus_share_cache(cpu, ctx->cpu);
374 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
375 rq->csd.func = __blk_mq_complete_request_remote;
378 smp_call_function_single_async(ctx->cpu, &rq->csd);
380 rq->q->softirq_done_fn(rq);
385 static void __blk_mq_complete_request(struct request *rq)
387 struct request_queue *q = rq->q;
389 if (!q->softirq_done_fn)
390 blk_mq_end_request(rq, rq->errors);
392 blk_mq_ipi_complete_request(rq);
396 * blk_mq_complete_request - end I/O on a request
397 * @rq: the request being processed
400 * Ends all I/O on a request. It does not handle partial completions.
401 * The actual completion happens out-of-order, through a IPI handler.
403 void blk_mq_complete_request(struct request *rq, int error)
405 struct request_queue *q = rq->q;
407 if (unlikely(blk_should_fake_timeout(q)))
409 if (!blk_mark_rq_complete(rq)) {
411 __blk_mq_complete_request(rq);
414 EXPORT_SYMBOL(blk_mq_complete_request);
416 int blk_mq_request_started(struct request *rq)
418 return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
420 EXPORT_SYMBOL_GPL(blk_mq_request_started);
422 void blk_mq_start_request(struct request *rq)
424 struct request_queue *q = rq->q;
426 trace_block_rq_issue(q, rq);
428 rq->resid_len = blk_rq_bytes(rq);
429 if (unlikely(blk_bidi_rq(rq)))
430 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
435 * Ensure that ->deadline is visible before set the started
436 * flag and clear the completed flag.
438 smp_mb__before_atomic();
441 * Mark us as started and clear complete. Complete might have been
442 * set if requeue raced with timeout, which then marked it as
443 * complete. So be sure to clear complete again when we start
444 * the request, otherwise we'll ignore the completion event.
446 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
447 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
448 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
449 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
451 if (q->dma_drain_size && blk_rq_bytes(rq)) {
453 * Make sure space for the drain appears. We know we can do
454 * this because max_hw_segments has been adjusted to be one
455 * fewer than the device can handle.
457 rq->nr_phys_segments++;
460 EXPORT_SYMBOL(blk_mq_start_request);
462 static void __blk_mq_requeue_request(struct request *rq)
464 struct request_queue *q = rq->q;
466 trace_block_rq_requeue(q, rq);
468 if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
469 if (q->dma_drain_size && blk_rq_bytes(rq))
470 rq->nr_phys_segments--;
474 void blk_mq_requeue_request(struct request *rq)
476 __blk_mq_requeue_request(rq);
478 BUG_ON(blk_queued_rq(rq));
479 blk_mq_add_to_requeue_list(rq, true);
481 EXPORT_SYMBOL(blk_mq_requeue_request);
483 static void blk_mq_requeue_work(struct work_struct *work)
485 struct request_queue *q =
486 container_of(work, struct request_queue, requeue_work.work);
488 struct request *rq, *next;
491 spin_lock_irqsave(&q->requeue_lock, flags);
492 list_splice_init(&q->requeue_list, &rq_list);
493 spin_unlock_irqrestore(&q->requeue_lock, flags);
495 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
496 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
499 rq->cmd_flags &= ~REQ_SOFTBARRIER;
500 list_del_init(&rq->queuelist);
501 blk_mq_insert_request(rq, true, false, false);
504 while (!list_empty(&rq_list)) {
505 rq = list_entry(rq_list.next, struct request, queuelist);
506 list_del_init(&rq->queuelist);
507 blk_mq_insert_request(rq, false, false, false);
511 * Use the start variant of queue running here, so that running
512 * the requeue work will kick stopped queues.
514 blk_mq_start_hw_queues(q);
517 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
519 struct request_queue *q = rq->q;
523 * We abuse this flag that is otherwise used by the I/O scheduler to
524 * request head insertation from the workqueue.
526 BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
528 spin_lock_irqsave(&q->requeue_lock, flags);
530 rq->cmd_flags |= REQ_SOFTBARRIER;
531 list_add(&rq->queuelist, &q->requeue_list);
533 list_add_tail(&rq->queuelist, &q->requeue_list);
535 spin_unlock_irqrestore(&q->requeue_lock, flags);
537 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
539 void blk_mq_cancel_requeue_work(struct request_queue *q)
541 cancel_delayed_work_sync(&q->requeue_work);
543 EXPORT_SYMBOL_GPL(blk_mq_cancel_requeue_work);
545 void blk_mq_kick_requeue_list(struct request_queue *q)
547 kblockd_schedule_delayed_work(&q->requeue_work, 0);
549 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
551 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
554 kblockd_schedule_delayed_work(&q->requeue_work,
555 msecs_to_jiffies(msecs));
557 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
559 void blk_mq_abort_requeue_list(struct request_queue *q)
564 spin_lock_irqsave(&q->requeue_lock, flags);
565 list_splice_init(&q->requeue_list, &rq_list);
566 spin_unlock_irqrestore(&q->requeue_lock, flags);
568 while (!list_empty(&rq_list)) {
571 rq = list_first_entry(&rq_list, struct request, queuelist);
572 list_del_init(&rq->queuelist);
574 blk_mq_end_request(rq, rq->errors);
577 EXPORT_SYMBOL(blk_mq_abort_requeue_list);
579 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
581 if (tag < tags->nr_tags) {
582 prefetch(tags->rqs[tag]);
583 return tags->rqs[tag];
588 EXPORT_SYMBOL(blk_mq_tag_to_rq);
590 struct blk_mq_timeout_data {
592 unsigned int next_set;
595 void blk_mq_rq_timed_out(struct request *req, bool reserved)
597 struct blk_mq_ops *ops = req->q->mq_ops;
598 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
601 * We know that complete is set at this point. If STARTED isn't set
602 * anymore, then the request isn't active and the "timeout" should
603 * just be ignored. This can happen due to the bitflag ordering.
604 * Timeout first checks if STARTED is set, and if it is, assumes
605 * the request is active. But if we race with completion, then
606 * we both flags will get cleared. So check here again, and ignore
607 * a timeout event with a request that isn't active.
609 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
613 ret = ops->timeout(req, reserved);
617 __blk_mq_complete_request(req);
619 case BLK_EH_RESET_TIMER:
621 blk_clear_rq_complete(req);
623 case BLK_EH_NOT_HANDLED:
626 printk(KERN_ERR "block: bad eh return: %d\n", ret);
631 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
632 struct request *rq, void *priv, bool reserved)
634 struct blk_mq_timeout_data *data = priv;
636 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
638 * If a request wasn't started before the queue was
639 * marked dying, kill it here or it'll go unnoticed.
641 if (unlikely(blk_queue_dying(rq->q))) {
643 blk_mq_end_request(rq, rq->errors);
648 if (time_after_eq(jiffies, rq->deadline)) {
649 if (!blk_mark_rq_complete(rq))
650 blk_mq_rq_timed_out(rq, reserved);
651 } else if (!data->next_set || time_after(data->next, rq->deadline)) {
652 data->next = rq->deadline;
657 static void blk_mq_timeout_work(struct work_struct *work)
659 struct request_queue *q =
660 container_of(work, struct request_queue, timeout_work);
661 struct blk_mq_timeout_data data = {
667 /* A deadlock might occur if a request is stuck requiring a
668 * timeout at the same time a queue freeze is waiting
669 * completion, since the timeout code would not be able to
670 * acquire the queue reference here.
672 * That's why we don't use blk_queue_enter here; instead, we use
673 * percpu_ref_tryget directly, because we need to be able to
674 * obtain a reference even in the short window between the queue
675 * starting to freeze, by dropping the first reference in
676 * blk_mq_freeze_queue_start, and the moment the last request is
677 * consumed, marked by the instant q_usage_counter reaches
680 if (!percpu_ref_tryget(&q->q_usage_counter))
683 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
686 data.next = blk_rq_timeout(round_jiffies_up(data.next));
687 mod_timer(&q->timeout, data.next);
689 struct blk_mq_hw_ctx *hctx;
691 queue_for_each_hw_ctx(q, hctx, i) {
692 /* the hctx may be unmapped, so check it here */
693 if (blk_mq_hw_queue_mapped(hctx))
694 blk_mq_tag_idle(hctx);
701 * Reverse check our software queue for entries that we could potentially
702 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
703 * too much time checking for merges.
705 static bool blk_mq_attempt_merge(struct request_queue *q,
706 struct blk_mq_ctx *ctx, struct bio *bio)
711 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
717 if (!blk_rq_merge_ok(rq, bio))
720 el_ret = blk_try_merge(rq, bio);
721 if (el_ret == ELEVATOR_BACK_MERGE) {
722 if (bio_attempt_back_merge(q, rq, bio)) {
727 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
728 if (bio_attempt_front_merge(q, rq, bio)) {
739 struct flush_busy_ctx_data {
740 struct blk_mq_hw_ctx *hctx;
741 struct list_head *list;
744 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
746 struct flush_busy_ctx_data *flush_data = data;
747 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
748 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
750 sbitmap_clear_bit(sb, bitnr);
751 spin_lock(&ctx->lock);
752 list_splice_tail_init(&ctx->rq_list, flush_data->list);
753 spin_unlock(&ctx->lock);
758 * Process software queues that have been marked busy, splicing them
759 * to the for-dispatch
761 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
763 struct flush_busy_ctx_data data = {
768 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
771 static inline unsigned int queued_to_index(unsigned int queued)
776 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
780 * Run this hardware queue, pulling any software queues mapped to it in.
781 * Note that this function currently has various problems around ordering
782 * of IO. In particular, we'd like FIFO behaviour on handling existing
783 * items on the hctx->dispatch list. Ignore that for now.
785 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
787 struct request_queue *q = hctx->queue;
790 LIST_HEAD(driver_list);
791 struct list_head *dptr;
794 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
797 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
798 cpu_online(hctx->next_cpu));
803 * Touch any software queue that has pending entries.
805 flush_busy_ctxs(hctx, &rq_list);
808 * If we have previous entries on our dispatch list, grab them
809 * and stuff them at the front for more fair dispatch.
811 if (!list_empty_careful(&hctx->dispatch)) {
812 spin_lock(&hctx->lock);
813 if (!list_empty(&hctx->dispatch))
814 list_splice_init(&hctx->dispatch, &rq_list);
815 spin_unlock(&hctx->lock);
819 * Start off with dptr being NULL, so we start the first request
820 * immediately, even if we have more pending.
825 * Now process all the entries, sending them to the driver.
828 while (!list_empty(&rq_list)) {
829 struct blk_mq_queue_data bd;
832 rq = list_first_entry(&rq_list, struct request, queuelist);
833 list_del_init(&rq->queuelist);
837 bd.last = list_empty(&rq_list);
839 ret = q->mq_ops->queue_rq(hctx, &bd);
841 case BLK_MQ_RQ_QUEUE_OK:
844 case BLK_MQ_RQ_QUEUE_BUSY:
845 list_add(&rq->queuelist, &rq_list);
846 __blk_mq_requeue_request(rq);
849 pr_err("blk-mq: bad return on queue: %d\n", ret);
850 case BLK_MQ_RQ_QUEUE_ERROR:
852 blk_mq_end_request(rq, rq->errors);
856 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
860 * We've done the first request. If we have more than 1
861 * left in the list, set dptr to defer issue.
863 if (!dptr && rq_list.next != rq_list.prev)
867 hctx->dispatched[queued_to_index(queued)]++;
870 * Any items that need requeuing? Stuff them into hctx->dispatch,
871 * that is where we will continue on next queue run.
873 if (!list_empty(&rq_list)) {
874 spin_lock(&hctx->lock);
875 list_splice(&rq_list, &hctx->dispatch);
876 spin_unlock(&hctx->lock);
878 * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
879 * it's possible the queue is stopped and restarted again
880 * before this. Queue restart will dispatch requests. And since
881 * requests in rq_list aren't added into hctx->dispatch yet,
882 * the requests in rq_list might get lost.
884 * blk_mq_run_hw_queue() already checks the STOPPED bit
886 blk_mq_run_hw_queue(hctx, true);
891 * It'd be great if the workqueue API had a way to pass
892 * in a mask and had some smarts for more clever placement.
893 * For now we just round-robin here, switching for every
894 * BLK_MQ_CPU_WORK_BATCH queued items.
896 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
898 if (hctx->queue->nr_hw_queues == 1)
899 return WORK_CPU_UNBOUND;
901 if (--hctx->next_cpu_batch <= 0) {
902 int cpu = hctx->next_cpu, next_cpu;
904 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
905 if (next_cpu >= nr_cpu_ids)
906 next_cpu = cpumask_first(hctx->cpumask);
908 hctx->next_cpu = next_cpu;
909 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
914 return hctx->next_cpu;
917 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
919 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state) ||
920 !blk_mq_hw_queue_mapped(hctx)))
925 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
926 __blk_mq_run_hw_queue(hctx);
934 kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work);
937 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
939 struct blk_mq_hw_ctx *hctx;
942 queue_for_each_hw_ctx(q, hctx, i) {
943 if ((!blk_mq_hctx_has_pending(hctx) &&
944 list_empty_careful(&hctx->dispatch)) ||
945 test_bit(BLK_MQ_S_STOPPED, &hctx->state))
948 blk_mq_run_hw_queue(hctx, async);
951 EXPORT_SYMBOL(blk_mq_run_hw_queues);
953 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
955 cancel_work(&hctx->run_work);
956 cancel_delayed_work(&hctx->delay_work);
957 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
959 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
961 void blk_mq_stop_hw_queues(struct request_queue *q)
963 struct blk_mq_hw_ctx *hctx;
966 queue_for_each_hw_ctx(q, hctx, i)
967 blk_mq_stop_hw_queue(hctx);
969 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
971 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
973 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
975 blk_mq_run_hw_queue(hctx, false);
977 EXPORT_SYMBOL(blk_mq_start_hw_queue);
979 void blk_mq_start_hw_queues(struct request_queue *q)
981 struct blk_mq_hw_ctx *hctx;
984 queue_for_each_hw_ctx(q, hctx, i)
985 blk_mq_start_hw_queue(hctx);
987 EXPORT_SYMBOL(blk_mq_start_hw_queues);
989 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
991 struct blk_mq_hw_ctx *hctx;
994 queue_for_each_hw_ctx(q, hctx, i) {
995 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
998 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
999 blk_mq_run_hw_queue(hctx, async);
1002 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1004 static void blk_mq_run_work_fn(struct work_struct *work)
1006 struct blk_mq_hw_ctx *hctx;
1008 hctx = container_of(work, struct blk_mq_hw_ctx, run_work);
1010 __blk_mq_run_hw_queue(hctx);
1013 static void blk_mq_delay_work_fn(struct work_struct *work)
1015 struct blk_mq_hw_ctx *hctx;
1017 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
1019 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
1020 __blk_mq_run_hw_queue(hctx);
1023 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1025 if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
1028 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1029 &hctx->delay_work, msecs_to_jiffies(msecs));
1031 EXPORT_SYMBOL(blk_mq_delay_queue);
1033 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1037 struct blk_mq_ctx *ctx = rq->mq_ctx;
1039 trace_block_rq_insert(hctx->queue, rq);
1042 list_add(&rq->queuelist, &ctx->rq_list);
1044 list_add_tail(&rq->queuelist, &ctx->rq_list);
1047 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
1048 struct request *rq, bool at_head)
1050 struct blk_mq_ctx *ctx = rq->mq_ctx;
1052 __blk_mq_insert_req_list(hctx, rq, at_head);
1053 blk_mq_hctx_mark_pending(hctx, ctx);
1056 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
1059 struct blk_mq_ctx *ctx = rq->mq_ctx;
1060 struct request_queue *q = rq->q;
1061 struct blk_mq_hw_ctx *hctx;
1063 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1065 spin_lock(&ctx->lock);
1066 __blk_mq_insert_request(hctx, rq, at_head);
1067 spin_unlock(&ctx->lock);
1070 blk_mq_run_hw_queue(hctx, async);
1073 static void blk_mq_insert_requests(struct request_queue *q,
1074 struct blk_mq_ctx *ctx,
1075 struct list_head *list,
1080 struct blk_mq_hw_ctx *hctx;
1082 trace_block_unplug(q, depth, !from_schedule);
1084 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1087 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1090 spin_lock(&ctx->lock);
1091 while (!list_empty(list)) {
1094 rq = list_first_entry(list, struct request, queuelist);
1095 BUG_ON(rq->mq_ctx != ctx);
1096 list_del_init(&rq->queuelist);
1097 __blk_mq_insert_req_list(hctx, rq, false);
1099 blk_mq_hctx_mark_pending(hctx, ctx);
1100 spin_unlock(&ctx->lock);
1102 blk_mq_run_hw_queue(hctx, from_schedule);
1105 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1107 struct request *rqa = container_of(a, struct request, queuelist);
1108 struct request *rqb = container_of(b, struct request, queuelist);
1110 return !(rqa->mq_ctx < rqb->mq_ctx ||
1111 (rqa->mq_ctx == rqb->mq_ctx &&
1112 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1115 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1117 struct blk_mq_ctx *this_ctx;
1118 struct request_queue *this_q;
1121 LIST_HEAD(ctx_list);
1124 list_splice_init(&plug->mq_list, &list);
1126 list_sort(NULL, &list, plug_ctx_cmp);
1132 while (!list_empty(&list)) {
1133 rq = list_entry_rq(list.next);
1134 list_del_init(&rq->queuelist);
1136 if (rq->mq_ctx != this_ctx) {
1138 blk_mq_insert_requests(this_q, this_ctx,
1143 this_ctx = rq->mq_ctx;
1149 list_add_tail(&rq->queuelist, &ctx_list);
1153 * If 'this_ctx' is set, we know we have entries to complete
1154 * on 'ctx_list'. Do those.
1157 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1162 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1164 init_request_from_bio(rq, bio);
1166 blk_account_io_start(rq, 1);
1169 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1171 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1172 !blk_queue_nomerges(hctx->queue);
1175 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1176 struct blk_mq_ctx *ctx,
1177 struct request *rq, struct bio *bio)
1179 if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) {
1180 blk_mq_bio_to_request(rq, bio);
1181 spin_lock(&ctx->lock);
1183 __blk_mq_insert_request(hctx, rq, false);
1184 spin_unlock(&ctx->lock);
1187 struct request_queue *q = hctx->queue;
1189 spin_lock(&ctx->lock);
1190 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1191 blk_mq_bio_to_request(rq, bio);
1195 spin_unlock(&ctx->lock);
1196 __blk_mq_free_request(hctx, ctx, rq);
1201 struct blk_map_ctx {
1202 struct blk_mq_hw_ctx *hctx;
1203 struct blk_mq_ctx *ctx;
1206 static struct request *blk_mq_map_request(struct request_queue *q,
1208 struct blk_map_ctx *data)
1210 struct blk_mq_hw_ctx *hctx;
1211 struct blk_mq_ctx *ctx;
1213 int op = bio_data_dir(bio);
1215 struct blk_mq_alloc_data alloc_data;
1217 blk_queue_enter_live(q);
1218 ctx = blk_mq_get_ctx(q);
1219 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1221 if (rw_is_sync(bio_op(bio), bio->bi_opf))
1222 op_flags |= REQ_SYNC;
1224 trace_block_getrq(q, bio, op);
1225 blk_mq_set_alloc_data(&alloc_data, q, BLK_MQ_REQ_NOWAIT, ctx, hctx);
1226 rq = __blk_mq_alloc_request(&alloc_data, op, op_flags);
1227 if (unlikely(!rq)) {
1228 __blk_mq_run_hw_queue(hctx);
1229 blk_mq_put_ctx(ctx);
1230 trace_block_sleeprq(q, bio, op);
1232 ctx = blk_mq_get_ctx(q);
1233 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1234 blk_mq_set_alloc_data(&alloc_data, q, 0, ctx, hctx);
1235 rq = __blk_mq_alloc_request(&alloc_data, op, op_flags);
1236 ctx = alloc_data.ctx;
1237 hctx = alloc_data.hctx;
1246 static int blk_mq_direct_issue_request(struct request *rq, blk_qc_t *cookie)
1249 struct request_queue *q = rq->q;
1250 struct blk_mq_hw_ctx *hctx = q->mq_ops->map_queue(q,
1252 struct blk_mq_queue_data bd = {
1257 blk_qc_t new_cookie = blk_tag_to_qc_t(rq->tag, hctx->queue_num);
1260 * For OK queue, we are done. For error, kill it. Any other
1261 * error (busy), just add it to our list as we previously
1264 ret = q->mq_ops->queue_rq(hctx, &bd);
1265 if (ret == BLK_MQ_RQ_QUEUE_OK) {
1266 *cookie = new_cookie;
1270 __blk_mq_requeue_request(rq);
1272 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1273 *cookie = BLK_QC_T_NONE;
1275 blk_mq_end_request(rq, rq->errors);
1283 * Multiple hardware queue variant. This will not use per-process plugs,
1284 * but will attempt to bypass the hctx queueing if we can go straight to
1285 * hardware for SYNC IO.
1287 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1289 const int is_sync = rw_is_sync(bio_op(bio), bio->bi_opf);
1290 const int is_flush_fua = bio->bi_opf & (REQ_PREFLUSH | REQ_FUA);
1291 struct blk_map_ctx data;
1293 unsigned int request_count = 0;
1294 struct blk_plug *plug;
1295 struct request *same_queue_rq = NULL;
1298 blk_queue_bounce(q, &bio);
1300 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1302 return BLK_QC_T_NONE;
1305 blk_queue_split(q, &bio, q->bio_split);
1307 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1308 blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1309 return BLK_QC_T_NONE;
1311 rq = blk_mq_map_request(q, bio, &data);
1313 return BLK_QC_T_NONE;
1315 cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num);
1317 if (unlikely(is_flush_fua)) {
1318 blk_mq_bio_to_request(rq, bio);
1319 blk_insert_flush(rq);
1323 plug = current->plug;
1325 * If the driver supports defer issued based on 'last', then
1326 * queue it up like normal since we can potentially save some
1329 if (((plug && !blk_queue_nomerges(q)) || is_sync) &&
1330 !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
1331 struct request *old_rq = NULL;
1333 blk_mq_bio_to_request(rq, bio);
1336 * We do limited pluging. If the bio can be merged, do that.
1337 * Otherwise the existing request in the plug list will be
1338 * issued. So the plug list will have one request at most
1342 * The plug list might get flushed before this. If that
1343 * happens, same_queue_rq is invalid and plug list is
1346 if (same_queue_rq && !list_empty(&plug->mq_list)) {
1347 old_rq = same_queue_rq;
1348 list_del_init(&old_rq->queuelist);
1350 list_add_tail(&rq->queuelist, &plug->mq_list);
1351 } else /* is_sync */
1353 blk_mq_put_ctx(data.ctx);
1356 if (!blk_mq_direct_issue_request(old_rq, &cookie))
1358 blk_mq_insert_request(old_rq, false, true, true);
1362 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1364 * For a SYNC request, send it to the hardware immediately. For
1365 * an ASYNC request, just ensure that we run it later on. The
1366 * latter allows for merging opportunities and more efficient
1370 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1372 blk_mq_put_ctx(data.ctx);
1378 * Single hardware queue variant. This will attempt to use any per-process
1379 * plug for merging and IO deferral.
1381 static blk_qc_t blk_sq_make_request(struct request_queue *q, struct bio *bio)
1383 const int is_sync = rw_is_sync(bio_op(bio), bio->bi_opf);
1384 const int is_flush_fua = bio->bi_opf & (REQ_PREFLUSH | REQ_FUA);
1385 struct blk_plug *plug;
1386 unsigned int request_count = 0;
1387 struct blk_map_ctx data;
1391 blk_queue_bounce(q, &bio);
1393 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1395 return BLK_QC_T_NONE;
1398 blk_queue_split(q, &bio, q->bio_split);
1400 if (!is_flush_fua && !blk_queue_nomerges(q)) {
1401 if (blk_attempt_plug_merge(q, bio, &request_count, NULL))
1402 return BLK_QC_T_NONE;
1404 request_count = blk_plug_queued_count(q);
1406 rq = blk_mq_map_request(q, bio, &data);
1408 return BLK_QC_T_NONE;
1410 cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num);
1412 if (unlikely(is_flush_fua)) {
1413 blk_mq_bio_to_request(rq, bio);
1414 blk_insert_flush(rq);
1419 * A task plug currently exists. Since this is completely lockless,
1420 * utilize that to temporarily store requests until the task is
1421 * either done or scheduled away.
1423 plug = current->plug;
1425 blk_mq_bio_to_request(rq, bio);
1427 trace_block_plug(q);
1429 blk_mq_put_ctx(data.ctx);
1431 if (request_count >= BLK_MAX_REQUEST_COUNT) {
1432 blk_flush_plug_list(plug, false);
1433 trace_block_plug(q);
1436 list_add_tail(&rq->queuelist, &plug->mq_list);
1440 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1442 * For a SYNC request, send it to the hardware immediately. For
1443 * an ASYNC request, just ensure that we run it later on. The
1444 * latter allows for merging opportunities and more efficient
1448 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1451 blk_mq_put_ctx(data.ctx);
1456 * Default mapping to a software queue, since we use one per CPU.
1458 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1460 return q->queue_hw_ctx[q->mq_map[cpu]];
1462 EXPORT_SYMBOL(blk_mq_map_queue);
1464 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1465 struct blk_mq_tags *tags, unsigned int hctx_idx)
1469 if (tags->rqs && set->ops->exit_request) {
1472 for (i = 0; i < tags->nr_tags; i++) {
1475 set->ops->exit_request(set->driver_data, tags->rqs[i],
1477 tags->rqs[i] = NULL;
1481 while (!list_empty(&tags->page_list)) {
1482 page = list_first_entry(&tags->page_list, struct page, lru);
1483 list_del_init(&page->lru);
1485 * Remove kmemleak object previously allocated in
1486 * blk_mq_init_rq_map().
1488 kmemleak_free(page_address(page));
1489 __free_pages(page, page->private);
1494 blk_mq_free_tags(tags);
1497 static size_t order_to_size(unsigned int order)
1499 return (size_t)PAGE_SIZE << order;
1502 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1503 unsigned int hctx_idx)
1505 struct blk_mq_tags *tags;
1506 unsigned int i, j, entries_per_page, max_order = 4;
1507 size_t rq_size, left;
1509 tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1511 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1515 INIT_LIST_HEAD(&tags->page_list);
1517 tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
1518 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1521 blk_mq_free_tags(tags);
1526 * rq_size is the size of the request plus driver payload, rounded
1527 * to the cacheline size
1529 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1531 left = rq_size * set->queue_depth;
1533 for (i = 0; i < set->queue_depth; ) {
1534 int this_order = max_order;
1539 while (this_order && left < order_to_size(this_order - 1))
1543 page = alloc_pages_node(set->numa_node,
1544 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1550 if (order_to_size(this_order) < rq_size)
1557 page->private = this_order;
1558 list_add_tail(&page->lru, &tags->page_list);
1560 p = page_address(page);
1562 * Allow kmemleak to scan these pages as they contain pointers
1563 * to additional allocations like via ops->init_request().
1565 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_KERNEL);
1566 entries_per_page = order_to_size(this_order) / rq_size;
1567 to_do = min(entries_per_page, set->queue_depth - i);
1568 left -= to_do * rq_size;
1569 for (j = 0; j < to_do; j++) {
1571 if (set->ops->init_request) {
1572 if (set->ops->init_request(set->driver_data,
1573 tags->rqs[i], hctx_idx, i,
1575 tags->rqs[i] = NULL;
1587 blk_mq_free_rq_map(set, tags, hctx_idx);
1592 * 'cpu' is going away. splice any existing rq_list entries from this
1593 * software queue to the hw queue dispatch list, and ensure that it
1596 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
1598 struct blk_mq_hw_ctx *hctx;
1599 struct blk_mq_ctx *ctx;
1602 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
1603 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
1605 spin_lock(&ctx->lock);
1606 if (!list_empty(&ctx->rq_list)) {
1607 list_splice_init(&ctx->rq_list, &tmp);
1608 blk_mq_hctx_clear_pending(hctx, ctx);
1610 spin_unlock(&ctx->lock);
1612 if (list_empty(&tmp))
1615 spin_lock(&hctx->lock);
1616 list_splice_tail_init(&tmp, &hctx->dispatch);
1617 spin_unlock(&hctx->lock);
1619 blk_mq_run_hw_queue(hctx, true);
1623 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
1625 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
1629 /* hctx->ctxs will be freed in queue's release handler */
1630 static void blk_mq_exit_hctx(struct request_queue *q,
1631 struct blk_mq_tag_set *set,
1632 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1634 unsigned flush_start_tag = set->queue_depth;
1636 blk_mq_tag_idle(hctx);
1638 if (set->ops->exit_request)
1639 set->ops->exit_request(set->driver_data,
1640 hctx->fq->flush_rq, hctx_idx,
1641 flush_start_tag + hctx_idx);
1643 if (set->ops->exit_hctx)
1644 set->ops->exit_hctx(hctx, hctx_idx);
1646 blk_mq_remove_cpuhp(hctx);
1647 blk_free_flush_queue(hctx->fq);
1648 sbitmap_free(&hctx->ctx_map);
1651 static void blk_mq_exit_hw_queues(struct request_queue *q,
1652 struct blk_mq_tag_set *set, int nr_queue)
1654 struct blk_mq_hw_ctx *hctx;
1657 queue_for_each_hw_ctx(q, hctx, i) {
1660 blk_mq_exit_hctx(q, set, hctx, i);
1664 static void blk_mq_free_hw_queues(struct request_queue *q,
1665 struct blk_mq_tag_set *set)
1667 struct blk_mq_hw_ctx *hctx;
1670 queue_for_each_hw_ctx(q, hctx, i)
1671 free_cpumask_var(hctx->cpumask);
1674 static int blk_mq_init_hctx(struct request_queue *q,
1675 struct blk_mq_tag_set *set,
1676 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1679 unsigned flush_start_tag = set->queue_depth;
1681 node = hctx->numa_node;
1682 if (node == NUMA_NO_NODE)
1683 node = hctx->numa_node = set->numa_node;
1685 INIT_WORK(&hctx->run_work, blk_mq_run_work_fn);
1686 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1687 spin_lock_init(&hctx->lock);
1688 INIT_LIST_HEAD(&hctx->dispatch);
1690 hctx->queue_num = hctx_idx;
1691 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
1693 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
1695 hctx->tags = set->tags[hctx_idx];
1698 * Allocate space for all possible cpus to avoid allocation at
1701 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1704 goto unregister_cpu_notifier;
1706 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
1712 if (set->ops->init_hctx &&
1713 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1716 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1720 if (set->ops->init_request &&
1721 set->ops->init_request(set->driver_data,
1722 hctx->fq->flush_rq, hctx_idx,
1723 flush_start_tag + hctx_idx, node))
1731 if (set->ops->exit_hctx)
1732 set->ops->exit_hctx(hctx, hctx_idx);
1734 sbitmap_free(&hctx->ctx_map);
1737 unregister_cpu_notifier:
1738 blk_mq_remove_cpuhp(hctx);
1742 static void blk_mq_init_cpu_queues(struct request_queue *q,
1743 unsigned int nr_hw_queues)
1747 for_each_possible_cpu(i) {
1748 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1749 struct blk_mq_hw_ctx *hctx;
1751 memset(__ctx, 0, sizeof(*__ctx));
1753 spin_lock_init(&__ctx->lock);
1754 INIT_LIST_HEAD(&__ctx->rq_list);
1757 /* If the cpu isn't online, the cpu is mapped to first hctx */
1761 hctx = q->mq_ops->map_queue(q, i);
1764 * Set local node, IFF we have more than one hw queue. If
1765 * not, we remain on the home node of the device
1767 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1768 hctx->numa_node = local_memory_node(cpu_to_node(i));
1772 static void blk_mq_map_swqueue(struct request_queue *q,
1773 const struct cpumask *online_mask)
1776 struct blk_mq_hw_ctx *hctx;
1777 struct blk_mq_ctx *ctx;
1778 struct blk_mq_tag_set *set = q->tag_set;
1781 * Avoid others reading imcomplete hctx->cpumask through sysfs
1783 mutex_lock(&q->sysfs_lock);
1785 queue_for_each_hw_ctx(q, hctx, i) {
1786 cpumask_clear(hctx->cpumask);
1791 * Map software to hardware queues
1793 for_each_possible_cpu(i) {
1794 /* If the cpu isn't online, the cpu is mapped to first hctx */
1795 if (!cpumask_test_cpu(i, online_mask))
1798 ctx = per_cpu_ptr(q->queue_ctx, i);
1799 hctx = q->mq_ops->map_queue(q, i);
1801 cpumask_set_cpu(i, hctx->cpumask);
1802 ctx->index_hw = hctx->nr_ctx;
1803 hctx->ctxs[hctx->nr_ctx++] = ctx;
1806 mutex_unlock(&q->sysfs_lock);
1808 queue_for_each_hw_ctx(q, hctx, i) {
1810 * If no software queues are mapped to this hardware queue,
1811 * disable it and free the request entries.
1813 if (!hctx->nr_ctx) {
1815 blk_mq_free_rq_map(set, set->tags[i], i);
1816 set->tags[i] = NULL;
1822 /* unmapped hw queue can be remapped after CPU topo changed */
1824 set->tags[i] = blk_mq_init_rq_map(set, i);
1825 hctx->tags = set->tags[i];
1826 WARN_ON(!hctx->tags);
1828 cpumask_copy(hctx->tags->cpumask, hctx->cpumask);
1830 * Set the map size to the number of mapped software queues.
1831 * This is more accurate and more efficient than looping
1832 * over all possibly mapped software queues.
1834 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
1837 * Initialize batch roundrobin counts
1839 hctx->next_cpu = cpumask_first(hctx->cpumask);
1840 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1844 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
1846 struct blk_mq_hw_ctx *hctx;
1849 queue_for_each_hw_ctx(q, hctx, i) {
1851 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1853 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1857 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared)
1859 struct request_queue *q;
1861 list_for_each_entry(q, &set->tag_list, tag_set_list) {
1862 blk_mq_freeze_queue(q);
1863 queue_set_hctx_shared(q, shared);
1864 blk_mq_unfreeze_queue(q);
1868 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1870 struct blk_mq_tag_set *set = q->tag_set;
1872 mutex_lock(&set->tag_list_lock);
1873 list_del_init(&q->tag_set_list);
1874 if (list_is_singular(&set->tag_list)) {
1875 /* just transitioned to unshared */
1876 set->flags &= ~BLK_MQ_F_TAG_SHARED;
1877 /* update existing queue */
1878 blk_mq_update_tag_set_depth(set, false);
1880 mutex_unlock(&set->tag_list_lock);
1883 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1884 struct request_queue *q)
1888 mutex_lock(&set->tag_list_lock);
1890 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
1891 if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) {
1892 set->flags |= BLK_MQ_F_TAG_SHARED;
1893 /* update existing queue */
1894 blk_mq_update_tag_set_depth(set, true);
1896 if (set->flags & BLK_MQ_F_TAG_SHARED)
1897 queue_set_hctx_shared(q, true);
1898 list_add_tail(&q->tag_set_list, &set->tag_list);
1900 mutex_unlock(&set->tag_list_lock);
1904 * It is the actual release handler for mq, but we do it from
1905 * request queue's release handler for avoiding use-after-free
1906 * and headache because q->mq_kobj shouldn't have been introduced,
1907 * but we can't group ctx/kctx kobj without it.
1909 void blk_mq_release(struct request_queue *q)
1911 struct blk_mq_hw_ctx *hctx;
1914 /* hctx kobj stays in hctx */
1915 queue_for_each_hw_ctx(q, hctx, i) {
1925 kfree(q->queue_hw_ctx);
1927 /* ctx kobj stays in queue_ctx */
1928 free_percpu(q->queue_ctx);
1931 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1933 struct request_queue *uninit_q, *q;
1935 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1937 return ERR_PTR(-ENOMEM);
1939 q = blk_mq_init_allocated_queue(set, uninit_q);
1941 blk_cleanup_queue(uninit_q);
1945 EXPORT_SYMBOL(blk_mq_init_queue);
1947 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
1948 struct request_queue *q)
1951 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
1953 blk_mq_sysfs_unregister(q);
1954 for (i = 0; i < set->nr_hw_queues; i++) {
1960 node = blk_mq_hw_queue_to_node(q->mq_map, i);
1961 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1966 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
1973 atomic_set(&hctxs[i]->nr_active, 0);
1974 hctxs[i]->numa_node = node;
1975 hctxs[i]->queue_num = i;
1977 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
1978 free_cpumask_var(hctxs[i]->cpumask);
1983 blk_mq_hctx_kobj_init(hctxs[i]);
1985 for (j = i; j < q->nr_hw_queues; j++) {
1986 struct blk_mq_hw_ctx *hctx = hctxs[j];
1990 blk_mq_free_rq_map(set, hctx->tags, j);
1991 set->tags[j] = NULL;
1993 blk_mq_exit_hctx(q, set, hctx, j);
1994 free_cpumask_var(hctx->cpumask);
1995 kobject_put(&hctx->kobj);
2002 q->nr_hw_queues = i;
2003 blk_mq_sysfs_register(q);
2006 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2007 struct request_queue *q)
2009 /* mark the queue as mq asap */
2010 q->mq_ops = set->ops;
2012 q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2016 q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2017 GFP_KERNEL, set->numa_node);
2018 if (!q->queue_hw_ctx)
2021 q->mq_map = blk_mq_make_queue_map(set);
2025 blk_mq_realloc_hw_ctxs(set, q);
2026 if (!q->nr_hw_queues)
2029 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2030 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2032 q->nr_queues = nr_cpu_ids;
2034 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2036 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2037 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2039 q->sg_reserved_size = INT_MAX;
2041 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2042 INIT_LIST_HEAD(&q->requeue_list);
2043 spin_lock_init(&q->requeue_lock);
2045 if (q->nr_hw_queues > 1)
2046 blk_queue_make_request(q, blk_mq_make_request);
2048 blk_queue_make_request(q, blk_sq_make_request);
2051 * Do this after blk_queue_make_request() overrides it...
2053 q->nr_requests = set->queue_depth;
2055 if (set->ops->complete)
2056 blk_queue_softirq_done(q, set->ops->complete);
2058 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2061 mutex_lock(&all_q_mutex);
2063 list_add_tail(&q->all_q_node, &all_q_list);
2064 blk_mq_add_queue_tag_set(set, q);
2065 blk_mq_map_swqueue(q, cpu_online_mask);
2067 mutex_unlock(&all_q_mutex);
2075 kfree(q->queue_hw_ctx);
2077 free_percpu(q->queue_ctx);
2080 return ERR_PTR(-ENOMEM);
2082 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2084 void blk_mq_free_queue(struct request_queue *q)
2086 struct blk_mq_tag_set *set = q->tag_set;
2088 mutex_lock(&all_q_mutex);
2089 list_del_init(&q->all_q_node);
2090 mutex_unlock(&all_q_mutex);
2092 blk_mq_del_queue_tag_set(q);
2094 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2095 blk_mq_free_hw_queues(q, set);
2098 /* Basically redo blk_mq_init_queue with queue frozen */
2099 static void blk_mq_queue_reinit(struct request_queue *q,
2100 const struct cpumask *online_mask)
2102 WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2104 blk_mq_sysfs_unregister(q);
2106 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues, online_mask);
2109 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2110 * we should change hctx numa_node according to new topology (this
2111 * involves free and re-allocate memory, worthy doing?)
2114 blk_mq_map_swqueue(q, online_mask);
2116 blk_mq_sysfs_register(q);
2120 * New online cpumask which is going to be set in this hotplug event.
2121 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2122 * one-by-one and dynamically allocating this could result in a failure.
2124 static struct cpumask cpuhp_online_new;
2126 static void blk_mq_queue_reinit_work(void)
2128 struct request_queue *q;
2130 mutex_lock(&all_q_mutex);
2132 * We need to freeze and reinit all existing queues. Freezing
2133 * involves synchronous wait for an RCU grace period and doing it
2134 * one by one may take a long time. Start freezing all queues in
2135 * one swoop and then wait for the completions so that freezing can
2136 * take place in parallel.
2138 list_for_each_entry(q, &all_q_list, all_q_node)
2139 blk_mq_freeze_queue_start(q);
2140 list_for_each_entry(q, &all_q_list, all_q_node) {
2141 blk_mq_freeze_queue_wait(q);
2144 * timeout handler can't touch hw queue during the
2147 del_timer_sync(&q->timeout);
2150 list_for_each_entry(q, &all_q_list, all_q_node)
2151 blk_mq_queue_reinit(q, &cpuhp_online_new);
2153 list_for_each_entry(q, &all_q_list, all_q_node)
2154 blk_mq_unfreeze_queue(q);
2156 mutex_unlock(&all_q_mutex);
2159 static int blk_mq_queue_reinit_dead(unsigned int cpu)
2161 cpumask_copy(&cpuhp_online_new, cpu_online_mask);
2162 blk_mq_queue_reinit_work();
2167 * Before hotadded cpu starts handling requests, new mappings must be
2168 * established. Otherwise, these requests in hw queue might never be
2171 * For example, there is a single hw queue (hctx) and two CPU queues (ctx0
2172 * for CPU0, and ctx1 for CPU1).
2174 * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list
2175 * and set bit0 in pending bitmap as ctx1->index_hw is still zero.
2177 * And then while running hw queue, flush_busy_ctxs() finds bit0 is set in
2178 * pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list.
2179 * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list
2182 static int blk_mq_queue_reinit_prepare(unsigned int cpu)
2184 cpumask_copy(&cpuhp_online_new, cpu_online_mask);
2185 cpumask_set_cpu(cpu, &cpuhp_online_new);
2186 blk_mq_queue_reinit_work();
2190 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2194 for (i = 0; i < set->nr_hw_queues; i++) {
2195 set->tags[i] = blk_mq_init_rq_map(set, i);
2204 blk_mq_free_rq_map(set, set->tags[i], i);
2210 * Allocate the request maps associated with this tag_set. Note that this
2211 * may reduce the depth asked for, if memory is tight. set->queue_depth
2212 * will be updated to reflect the allocated depth.
2214 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2219 depth = set->queue_depth;
2221 err = __blk_mq_alloc_rq_maps(set);
2225 set->queue_depth >>= 1;
2226 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2230 } while (set->queue_depth);
2232 if (!set->queue_depth || err) {
2233 pr_err("blk-mq: failed to allocate request map\n");
2237 if (depth != set->queue_depth)
2238 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2239 depth, set->queue_depth);
2244 struct cpumask *blk_mq_tags_cpumask(struct blk_mq_tags *tags)
2246 return tags->cpumask;
2248 EXPORT_SYMBOL_GPL(blk_mq_tags_cpumask);
2251 * Alloc a tag set to be associated with one or more request queues.
2252 * May fail with EINVAL for various error conditions. May adjust the
2253 * requested depth down, if if it too large. In that case, the set
2254 * value will be stored in set->queue_depth.
2256 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2258 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2260 if (!set->nr_hw_queues)
2262 if (!set->queue_depth)
2264 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2267 if (!set->ops->queue_rq || !set->ops->map_queue)
2270 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2271 pr_info("blk-mq: reduced tag depth to %u\n",
2273 set->queue_depth = BLK_MQ_MAX_DEPTH;
2277 * If a crashdump is active, then we are potentially in a very
2278 * memory constrained environment. Limit us to 1 queue and
2279 * 64 tags to prevent using too much memory.
2281 if (is_kdump_kernel()) {
2282 set->nr_hw_queues = 1;
2283 set->queue_depth = min(64U, set->queue_depth);
2286 * There is no use for more h/w queues than cpus.
2288 if (set->nr_hw_queues > nr_cpu_ids)
2289 set->nr_hw_queues = nr_cpu_ids;
2291 set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2292 GFP_KERNEL, set->numa_node);
2296 if (blk_mq_alloc_rq_maps(set))
2299 mutex_init(&set->tag_list_lock);
2300 INIT_LIST_HEAD(&set->tag_list);
2308 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2310 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2314 for (i = 0; i < nr_cpu_ids; i++) {
2316 blk_mq_free_rq_map(set, set->tags[i], i);
2322 EXPORT_SYMBOL(blk_mq_free_tag_set);
2324 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2326 struct blk_mq_tag_set *set = q->tag_set;
2327 struct blk_mq_hw_ctx *hctx;
2330 if (!set || nr > set->queue_depth)
2334 queue_for_each_hw_ctx(q, hctx, i) {
2337 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2343 q->nr_requests = nr;
2348 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2350 struct request_queue *q;
2352 if (nr_hw_queues > nr_cpu_ids)
2353 nr_hw_queues = nr_cpu_ids;
2354 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2357 list_for_each_entry(q, &set->tag_list, tag_set_list)
2358 blk_mq_freeze_queue(q);
2360 set->nr_hw_queues = nr_hw_queues;
2361 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2362 blk_mq_realloc_hw_ctxs(set, q);
2364 if (q->nr_hw_queues > 1)
2365 blk_queue_make_request(q, blk_mq_make_request);
2367 blk_queue_make_request(q, blk_sq_make_request);
2369 blk_mq_queue_reinit(q, cpu_online_mask);
2372 list_for_each_entry(q, &set->tag_list, tag_set_list)
2373 blk_mq_unfreeze_queue(q);
2375 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2377 void blk_mq_disable_hotplug(void)
2379 mutex_lock(&all_q_mutex);
2382 void blk_mq_enable_hotplug(void)
2384 mutex_unlock(&all_q_mutex);
2387 static int __init blk_mq_init(void)
2389 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
2390 blk_mq_hctx_notify_dead);
2392 cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE, "block/mq:prepare",
2393 blk_mq_queue_reinit_prepare,
2394 blk_mq_queue_reinit_dead);
2397 subsys_initcall(blk_mq_init);